Research On Multi-Kernel Learning And Graph Regularization Based Cross-modal Hashing Retrieval

With the development of information technology,the era of big data has arrived,and the amount of multimedia data on the Internet grows explosively.Multi-modality is an important feature of big data.In order to deal with the problem of retrieving big data,cross-modal retrieval based on hashing been an active topic in computer vision research.The multi-modality feature of big data refers to that data have many modals,such as text and image,and cross-modal hashing methods map the heterogeneous data features of multiple modals into unified hash codes via hash mapping functions,and consequently achieving cross-modal retrieval.Hashing methods are able to not only speed up the retrieval speed,but also reduce the consumption of storage space.This paper studies the application of multi-kernel learning theory in the problem of cross modal hashing,and proposes two cross modal hashing methods,namely,multi-kernel supervised cross-modal hashing(MKSCMH)and multi-kernel supervised hashing with graph regularization for cross-modal retrieval(MKSRH).Compared with some recently proposed algorithms,experiments on the Wiki dataset and the NUS-WIDE dataset show the effectiveness of the proposed algorithm.The main contributions of this dissertation are summarized as follows:(1)An algorithm of multi-kernel supervised cross-modal hashing(MKSCMH)is proposed.Because the retrieved data has the characteristics of heterogeneous and irregular,large sample scale and uneven sample distribution,it is difficult to fully utilize the data information by using a single kernel function for nonlinear mapping.Therefore,the multi-kernel learning method is used to map the data to high-dimensional space with the combination of multiple kernel functions,and the AdaBoost algorithm is used to learn multiple classification hyperplanes in kernel space.The obtained classification hyperplanes are used as hash mapping functions,and then the hash codes are produced by symbolic functions.(2)An algorithm of multi-kernel supervised hashing with graph regularization for cross-modal retrieval(MKSRH)is proposed.Using the method of multi-kernel learning,the data is mapped to the high dimensional kernel space,and the AdaBoost algorithm is used to learn multiple classification hyperplanes in high dimensional space.However,these classification hyperplanes cannot guarantee the optimal classification results.Consequently,we use graph regularizatlion to fine tune the original hash codes obtained by mapping,and use support vector machine to tain classifiers to achieve classification hyperplanes refinement.By optimizing the classification hyperplanes,the obtained hash code can achieve higher retrieval precision.